Jev for Educators: Formative Only, Loudly

Say it before anyone asks: this page is not about letting a model grade your students. AI for teachers has earned its bad reputation by promising exactly that, and the honest version is smaller and more useful. Jev, TypeSafe AI's decision model, answers closed questions about student work fast and cheaply, which makes it good at practice feedback, question analysis, and sorting. It is not good at deciding anyone's future, and you shouldn't let it try.

The policy line between formative and summative (and why it holds) belongs to the AI grading guide. This page is the teacher's workflow on the permitted side of that line.

Classroom feedback automation for practice work

The formative loop is where the math actually works for a teacher. A practice set of 30 students times 10 short-answer questions is 300 reads, and the useful feedback on each is usually one of a handful of things: correct, correct but missing a step, a specific misconception, or off-topic.

That's a closed question per response. Write the choices as the misconceptions you already know your students have ("confused area with perimeter", "sign error in step two", "correct method, arithmetic slip") and each answer comes back labeled. The student sees targeted feedback on submission and retries while the idea is still warm. You see which label dominated.

Keep it to work where a wrong verdict costs a retry, not a mark: practice sets, exit tickets, draft checks, homework that isn't graded for credit. That's the whole scope.

Item analysis: the under-hyped win

The best use isn't feedback to students at all; it's feedback to you. Item analysis at full coverage answers the questions every teacher asks after a quiz: which question did the class miss, and which wrong answer did they converge on? A misconception shared by twelve students is a re-teach agenda for tomorrow, not twelve separate red marks.

One cataloged build points the same machine at exam prep from the student side: a builder gave Jev 80 real exam questions and 297 practice ones, and in 80 seconds it ranked which were most likely to appear on the real exam, for $0.0256 (build, as reported). Whether the predictions held isn't in the receipt, so treat it as a demonstration of cost and speed, not accuracy. The teacher-side version is sharper: map your practice bank against your curriculum objectives and find the objectives nothing practices.

AI for teachers: what teachers keep, loudly

Every mark that counts toward a grade, a transcript, a placement, or a report to parents is a teacher decision. Full stop. Verdicts can pre-sort and flag, with the evidence attached, but the teacher reads and decides, and the student can always ask a person why.

The reasons are practical, not sentimental. Model judges reward fluent writing over correct thinking unless you test for it deliberately. The same answer can get a slightly different probability on a different day, and students are owed consistent marks. And students probe graders the way spammers probe filters.

Audit posture, even for practice work:

• Log the question, its version, and the verdict for every piece of feedback a student sees.
• Spot-check a sample by hand every week, and check whether any group of students is flagged differently than you'd expect.
• Let students contest any automated feedback and reach a teacher.

On student data: what you're allowed to send to an outside service is set by your school, your district, and the law where you teach. Check with whoever owns that before the first run, send the minimum (the answer text, not the student's name), and read the vendor terms at docs.typesafe.ai. This is not legal advice, and we mean that.

What it can't do in a classroom

Jev doesn't write. It can't explain a concept, write a model answer, or chat with a student who's stuck; that's a chat model or, better, you. It can't score long essays the way a careful reader does. And it doesn't know your students. It knows the question you asked and the text you sent, which is precisely why the teacher stays in charge.

Frequently asked questions

Can AI grade my students' work?

It can pre-sort practice work and flag likely misconceptions; grades that count stay with the teacher. The grading guide draws that line in detail.

What is classroom feedback automation?

Instant, specific feedback on low-stakes practice, where each response gets labeled with the misconception it shows so students can retry immediately. The teacher sees the pattern across the class.

What does it cost for a class?

A class set of practice responses runs to fractions of a cent per verdict in builder-reported workloads; see the cost table for sourced numbers.

Is it safe to send student work to an AI service?

Only within your school's and jurisdiction's rules, with the minimum data sent and no names attached. Check policy first; this is not legal advice.

Where should a teacher start?

With item analysis on a quiz you've already graded by hand, so you can compare the model's misconception labels to your own before any student ever sees one.